Key points
- OpenMAIC is Tsinghua University's open-source multi-agent interactive classroom (MAIC) platform, flipping "MOOC = one video for N students" into "MAIC = N agents serving 1 student." Upload a PDF or describe a topic → a full AI classroom in ~30 minutes, for under $2.
- Timeline: MAIC launched on China's national smart education platform in April 2025; open-sourced on GitHub (THU-MAIC/OpenMAIC) on 2026-03-16; rapid releases added Ollama support (v0.1.1), Deep Interactive Mode (v0.2.0), VoxCPM2 TTS + new models (v0.2.1), and an AMD ROCm partnership (May 2026).
- Architecture: three layers — Next.js frontend, Core Engine (two-stage Generation Pipeline: outline → scenes; LangGraph orchestration), and a provider layer supporting OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Kimi, MiniMax, Grok, Ollama, Lemonade, and more.
- Cloud: AMD Instinct GPU + ROCm for compute-heavy course generation.
- Edge: AMD Ryzen AI PC (GPU+NPU) via the Lemonade OpenAI-compatible local server, for low-latency multi-agent dialogue, TTS/ASR, privacy, and offline use.
- OpenClaw skill (
clawhub install openmaic): generate classes from 20+ chat platforms (Slack, Discord, Telegram, Feishu); hosted at openmaic.chat or self-hosted. - Validated in real classrooms: 700+ students, 100K+ learning records, 240K+ visits, 84.1% satisfaction.
- Academic backing: *From MOOC to MAIC* (JCST 2026, DOI: 10.1007/s11390-025-6000-0), Slide2Lecture, EduCraft, LongWriter-V, SimClass (NAACL 2025).
- GitHub: https://github.com/THU-MAIC/OpenMAIC
- Live demo: https://openmaic.chat/
- AMD blog: https://www.amd.com/en/blogs/2026/reimagining-ai-native-education-through-multi-agent-interactive-classroom-on-amd-rocm.html
- Paper: JCST 2026, DOI: 10.1007/s11390-025-6000-0
Agent roles
| Agent | Role | |-------|------| | Lesson Planner | Analyzes documents, builds outline and learning path | | Content Generator | Slides, quizzes, simulations, voice scripts | | AI Teacher | Lectures, analogies, whiteboard drawing, voice | | AI Teaching Assistant | Explanations, quiz management, detects struggling students | | AI Classmates | 4 archetypes: ask questions, debate, offer perspectives | | Director Agent | Controls turn-taking, pacing, transitions |
28+ action types (speak, draw, quiz, highlight, laser pointer…) and four classroom modes: discussion, roundtable debate, collaborative Q&A, whiteboard interaction.
Deep Interactive Mode (v0.2.0)
Students go from passive viewing to hands-on: 3D visualization, parameterized simulations, knowledge mini-games, mind maps, and in-browser coding. The AI teacher can actively manipulate the UI — highlighting regions, setting conditions, giving hints.
Edge-cloud deployment with AMD
Integrations & validation
Commentary
1. Social dynamics beats content generation — the LangGraph state machine simulating classmates raising hands and debates is the hard-to-copy differentiator. 2. Cost structure shifted — sub-$2, 30-minute course generation is an order-of-magnitude change in content supply. 3. Edge-cloud split matches workloads — heavy generation in the cloud, latency-sensitive interaction at the edge. 4. AGPL-3.0 signals education infrastructure should remain public rather than privatized.